Yoga is effective for mild to moderate chronic low back pain (cLBP), but its comparative effectiveness with physical therapy (PT) is unknown. Moreover, little is known about yoga’s effectiveness in underserved patients with more severe functional disability and pain.
One-sided 95% lower confidence limits were 0.83 (RMDQ) and 0.97 (pain), demonstrating noninferiority of yoga to PT. However, yoga was not superior to education for either outcome. Yoga and PT were similar for most secondary outcomes. Yoga and PT participants were 21 and 22 percentage points less likely, respectively, than education participants to use pain medication at 12 weeks. Improvements in yoga and PT groups were maintained at 1 year with no differences between maintenance strategies. Frequency of adverse events, mostly mild self-limited joint and back pain, did not differ between the yoga and PT groups.
Low back pain is the leading cause of disability globally (1). Total annual back pain–related costs in the United States exceed $200 billion (2). Chronic low back pain (cLBP) affects approximately 10% of U.S. adults (3), but overall patient satisfaction with cLBP treatment is low (4). The impact of cLBP is greater in racial or ethnic minorities and in people of lower socioeconomic status (SES) (5). Physical therapy (PT), comprising individually tailored stretching and strengthening exercises, is the most common evidence-based, reimbursable, and nonpharmacologic physician referral for cLBP (6, 7). Clinical guidelines (8, 9), meta-analyses (10), and several large randomized controlled trials (11–13) also support yoga, a practice including physical poses, breathing exercises, and meditation, as an effective cLBP treatment.
To improve cLBP care, physicians, patients, and payers need to know how novel therapies like yoga compare with established treatments like PT. Noninferiority trials determine whether a new therapy is statistically as effective as an accepted treatment (14). This is particularly useful when the new therapy may have other potential benefits, such as lower cost. The U.S. Food and Drug Administration uses noninferiority as one criterion for approving new pharmaceuticals (15). Although no criteria have been established for when a new nonpharmacologic therapy should be integrated into mainstream clinical practice, demonstrating noninferiority to effective, reimbursed, and nonpharmacologic treatments is one reasonable requirement. Thus, we conducted a randomized controlled comparative effectiveness trial testing whether yoga was noninferior to PT in adults with cLBP.
The study protocol (16) and treatment manuals (17–20) were published previously. We conducted a 52-week, assessor-blinded randomized trial of yoga, PT, and education (a self-care book and newsletters) for adults with nonspecific cLBP. The study was advertised as a comparison of 3 credible cLBP treatments. All participants had access to usual medical care. The study consisted of a 12-week treatment phase and 40-week maintenance phase. The primary hypothesis was that yoga is noninferior to PT in the treatment phase for improving both back-related function and pain intensity. The secondary hypothesis was that both yoga and PT are superior to education for improving function and pain.
The maintenance phase compared the effectiveness of different strategies for ongoing yoga and PT. Yoga participants who completed 1 or more yoga classes in the treatment phase were randomly assigned at 12 weeks to yoga drop-in classes or home practice. Physical therapy patients who completed 1 or more PT appointments in the treatment phase were randomly assigned to PT booster sessions or home practice. We hypothesized that at 52 weeks, yoga drop-in classes are superior to yoga home practice and PT booster sessions are superior to PT home practice.
Participants were recruited from June 2012 to November 2013. Follow-up was completed in November 2014. The original research protocol and a summary of protocol changes are in Supplement 1.
Setting and Participants
The study was done at a large academic safety-net hospital and 7 affiliated, federally qualified community health centers located in diverse neighborhoods. Staff doing data collection, entry, and analysis were masked to treatment assignment. The Boston University Institutional Review Board approved the study before data collection.
We enrolled English-speaking adults aged 18 to 64 years who reported nonspecific low back pain lasting at least 12 weeks with an average pain intensity in the previous week of 4 or greater on an 11-point (0 to 10) numerical rating scale. Persons with specific causes of cLBP (for example, spinal stenosis) were excluded. Eligibility criteria are in Table 1 of Supplement 2. Recruitment strategies included clinician referrals, mailing letters to patients with cLBP who were identified through electronic health records, and distributing flyers in clinics and surrounding neighborhoods.
Randomization and Interventions
After initial telephone screening, staff confirmed eligibility and obtained written informed consent during in-person meetings. Staff entered participants into StudyTRAX (ScienceTrax), a data management platform. StudyTRAX generated a randomization sequence using permuted block randomization with varying block sizes and a 2:2:1 ratio of yoga, PT, and education. After participants completed baseline surveys, unmasked staff informed them of their treatment assignments. Enrollment and randomization proceeded in 4 sequential cohorts of approximately 80 participants each.
Yoga participants who attended at least 1 class in the treatment phase continued into the maintenance phase and were randomly assigned to weekly drop-in yoga classes or home practice only. Physical therapy participants with at least 1 PT visit in the treatment phase were randomly assigned to attend 5 booster sessions or home practice only during the maintenance phase. Education participants continued into the maintenance phase without additional randomization.
A manualized yoga protocol (18) of 12 weekly 75-minute classes was adapted from previous studies in similar populations (21, 22) with input from expert yoga instructors, investigators, and former study participants (Table 2 of Supplement 2). Thirteen yoga instructors completed 8 hours of training and taught classes at 6 sites. Instructor assignments ensured a participant–instructor ratio of less than 5:1. Each class began with relaxation and meditation exercises, yoga breathing, and yoga philosophy. It continued with yoga poses and concluded with relaxation. Pose variations and aids (such as chair, strap, and blocks) accommodated various abilities. Thirty minutes of daily home practice, facilitated by a DVD, a manual, and take-home yoga supplies, was strongly encouraged. Participants recorded time spent practicing. Staff observed approximately 10% of classes to assess protocol fidelity by using a checklist. Maintenance phase classes were similarly structured except for a higher participant–instructor ratio (approximately 8:1).
The manualized PT protocol (Figure 1 and Table 3 of Supplement 2) incorporated treatment-based classification (23, 24), graded exercise (25), and screening for fear-avoidance beliefs (26). Eight physical therapists delivered the intervention in 1 hospital-based and 2 community-based PT clinics. Physical therapists completed 8 hours of in-person training and Web-based modules (27). Participants were advised to attend fifteen 60-minute appointments over 12 weeks. Appointments included one-on-one work with the therapist and supervised aerobic exercise. All participants completed the Fear-Avoidance Beliefs Questionnaire (28). For PT participants with a high fear-avoidance score (≥29 on the work subscale), therapists provided The Back Book (29) and reinforced its psychologically informed principles to lower fear avoidance. Participants received written instructions and supplies for home practice and logged the number of exercises completed daily. Staff assessed protocol fidelity by reviewing therapists’ treatment flowsheets. Participants randomly assigned to booster sessions during the maintenance phase were advised to see the therapist at 4, 6, 8, 10, and 12 months.
Education participants received The Back Pain Helpbook (30), which includes information on cLBP self-management, stretching, strengthening, and the role of emotions and fear avoidance. Previous cLBP trials (11, 12, 31) used this book as a credible control intervention. We provided a recommended reading schedule (Table 4 of Supplement 2). Every 3 weeks, participants received a 1- to 2-page newsletter (32), summarizing main points from assigned chapters, and a 5- to 10-minute check-in call from staff. In the maintenance phase, we made brief check-in calls every 6 weeks to encourage continued review of the book.
Outcomes and Follow-up
Coprimary outcomes were change from baseline to 12 weeks in scores on the modified Roland Morris Disability Questionnaire (RMDQ), a 23-point measure of back-related function with higher scores indicating worse function (33), and in pain (using an 11-point numerical rating scale for average intensity in the previous week, where 0 indicated no pain and 10 indicated worst pain possible) (34). Secondary outcomes included self-reported pain medication use in the previous week (yes or no), global improvement (7-point scale from extremely worsened to extremely improved), patient satisfaction with interventions (5-point scale from very dissatisfied to very satisfied) (35), and health-related quality of life (Short Form-36 Health Survey) (36). We collected data on work productivity, a secondary outcome, and will report them separately in a cost-effectiveness analysis. We also plan to report data on other exploratory measures collected (such as fear-avoidance beliefs, pain self-efficacy, depression, anxiety, and sleep) separately. Attendance at all yoga and PT sessions was recorded. Participants attending at least 9 yoga or 11 PT sessions (≥75% or ≥73% of sessions, respectively) were defined as adherent a priori. Education participants were asked how much of the book they had read. Those reporting having read at least 75% were defined as adherent a priori.
After baseline data collection, study staff that was blinded to treatment assignment collected paper surveys completed at 6, 12, 26, 40, and 52 weeks. Adverse events were elicited directly from participants and in surveys. Participants received $100 gift cards after completing surveys at 12 and 52 weeks and $50 gift cards after the baseline questionnaire and surveys at 6, 26, and 40 weeks.
The study was designed and powered to detect if yoga was noninferior to PT at 12 weeks for both primary outcomes. Noninferiority margins were prespecified for RMDQ (1.5) and pain (1.0) by halving the minimal clinically important difference (37). Although some controversy exists about the minimal clinically important difference for RMDQ (38, 39) and back pain intensity (40, 41), 3.0 and 2.0, respectively, are reasonable and commonly used. Because we required noninferiority for both primary outcomes, adjustment for multiple testing was unnecessary. Assuming 20% attrition and previously published variances (11, 12), the target sample size of 320 provided 81% and 90% power to detect noninferiority at 12 weeks of yoga to PT for function and pain, respectively.
We used analysis of variance and chi-square tests to assess between-group differences in baseline variables. For coprimary outcomes during the treatment phase, we did 1-sided 2-sample t tests to determine whether yoga was noninferior to PT for change from baseline to 12 weeks. We controlled for potential confounders, defined a priori by a baseline imbalance between groups (P < 0.10), by using multiple linear regression. To comply with the journal editors’ recommendations for handling missing data in our primary outcome analyses, we present findings based on multiple imputation using regression modeling in SAS PROC MI. Variables used in the imputation were treatment group, baseline values, and week-6 scores. Any missing week-6 values were imputed before imputation for week-12 values. Ten imputed data sets were created. Using these, the analysis for back pain score was unadjusted, whereas the analysis of RMDQ was adjusted for baseline RMDQ. Analyses using the last observation carried forward, our prespecified approach for handling missing values, are presented in Supplement 2.
For secondary outcomes, we used the last observation carried forward to manage missing data and did not adjust for multiple testing. We did 2-sided 2-sample t tests to determine whether yoga and PT were superior to education.
We also performed responder analyses comparing the proportion of participants in each group with clinically meaningful change (≥30% decrease from baseline to 12 weeks) (40). Self-reported pain medication use at 12 weeks (any vs. none) and medication subtypes (nonsteroidal anti-inflammatory drugs, acetaminophen, or opioids) were examined using logistic regression adjusted for baseline use. Global improvement and patient satisfaction were compared using multiple logistic regression. Health-related quality of life was compared using multiple linear regression. Per protocol analyses of the treatment phase included participants meeting a priori definitions of adherence. Adverse events were compared using the Fisher exact test.
For the maintenance phase, we did longitudinal analyses with a 5-part treatment variable (yoga drop-in classes, yoga home practice, PT booster sessions, PT home practice, and education), incorporating all RMDQ and pain measurements from weeks 12 to 52 (42). The general linear model for correlated data using SAS PROC MIXED with a repeated statement was used with an unstructured covariance and a treatment-by-time interaction examining differences in patterns over time. Specific hypotheses were tested using contrasts. Missing data points were not replaced.
All analyses were performed with SAS, version 9.3 (SAS Institute).
Role of the Funding Source
This study was funded by the National Center for Complementary and Integrative Health. This agency had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, and approval of the manuscript; or decision to submit the manuscript for publication.
From June 2012 to November 2013, we screened 1663 people. Of these, 479 (29%) met eligibility criteria and 320 (19%) were enrolled and randomly assigned to yoga, PT, or education (Figure 2 and Table 5 of Supplement 2). Most study participants were women, were nonwhite, were not college graduates, and were earning $30 000 or less annually (Table 1; Table 6 of Supplement 2).
On average, participants reported moderate to severe functional impairment and pain. More than two thirds used analgesics for back pain. Baseline mean between-group differences were present for RMDQ, sex, and body mass index (P = 0.032, 0.088, and 0.099, respectively). However, only baseline RMDQ was identified as a confounder for the RMDQ analyses.
Adherence to Interventions and Loss to Follow-up
During the treatment phase, median yoga attendance was 7 classes (interquartile range, 3 to 10). Median PT attendance was 7 appointments (interquartile range, 2 to 12). Home practice was reported by 95 yoga participants (75%) and 83 PT participants (64%). Of these, yoga participants practiced a median of 27 minutes (interquartile range, 17 to 35 minutes) 4 days per week. Physical therapy participants did a median of 4 exercises (interquartile range, 3 to 5) 4 days per week. Fewer than half of the participants met adherence criteria: 56 yoga (44%), 46 PT (36%), and 28 education (44%) participants. Of 59 participants randomly assigned to yoga drop-in classes during the maintenance phase, 31 (53%) attended at least 1 drop-in class (median, 13; interquartile range, 4 to 22). Of 54 participants randomly assigned to PT booster sessions, 30 (56%) attended at least 1 appointment (median, 2; interquartile range, 1 to 3). Follow-up was lower in PT than in yoga or education at 12 weeks (88% vs. 98% and 95%, respectively) and 52 weeks (84% vs. 93% and 93%, respectively).
Improvement in RMDQ for yoga (mean within-group change, −3.8 [95% CI, −4.6 to −2.9]) was noninferior to that for PT (mean within-group change, −3.5 [CI, −4.5 to −2.6]) (Table 2 and Figure). The mean difference in RMDQ between yoga and PT was −0.26 (1-sided CI, −∞ to 0.83).
Decreased pain for yoga (mean within-group change, −1.7 [CI, −2.1 to −1.4]) was noninferior to that for PT (mean within-group change, −2.3 [CI, −2.7 to −1.9]). The mean difference in pain between yoga and PT was 0.51 (1-sided CI, −∞ to 0.97). Noninferiority plots for primary outcomes are shown in Figure 3 of Supplement 2. Analyses using the last observation carried forward to account for missing data yielded similar results (Table 7 of Supplement 2).
Secondary and Exploratory Outcomes
Yoga and PT were not superior to education at 12 weeks for RMDQ (Table 2). However, both yoga and PT were more likely than education to have clinically meaningful responses in RMDQ (Table 3). Forty-eight percent of yoga participants versus 23% of education participants responded (odds ratio, 3.1 [CI, 1.6 to 6.2]). Thirty-seven percent of PT participants responded (odds ratio [vs. education], 2.0 [CI, 1.0 to 4.0]). For pain, yoga was not superior to education (mean between-group difference, −0.33 [CI, −0.97 to 0.32]) but PT was (mean between-group difference, −0.84 [CI, −1.5 to −0.18]). The only significant between-group difference in clinically meaningful response in pain was in the PT group compared with education (43% vs. 25%; odds ratio, 2.3 [CI, 1.1 to 4.5]).
At 12 weeks, yoga and PT participants were 21 and 22 percentage points, respectively, less likely than education participants to use any pain medication (Table 3). Although PT participants were less likely than education participants to use acetaminophen, there were no other significant differences in medication subgroups. Self-rated global improvement and satisfaction with the intervention did not significantly differ between yoga and PT. Global improvement for PT, but not for yoga, was superior to that for education. Satisfaction with yoga and PT were both superior to that with education. No significant between-group differences were seen in Short Form-36 Health Survey scores.
Given the low proportion of participants meeting prespecified adherence criteria (36% to 44%), per protocol between-group comparisons are highly susceptible to bias and are not presented. However, mean within-group RMDQ changes for adherent participants at 12 weeks were −4.6, −5.7, and −2.7 for yoga, PT, and education, respectively (Table 8 of Supplement 2). Clinically meaningful improvement in RMDQ scores occurred in 57%, 56%, and 21% of participants, respectively. For pain intensity at 12 weeks, mean within-group changes for adherent participants were −2.1, −2.6, and −1.3 for yoga, PT, and education, respectively. Clinically meaningful improvement in pain intensity occurred in 50%, 52%, and 14% of participants, respectively.
Adverse events, mostly mild self-limited joint and back pain, were reported in 9 yoga, 14 PT, and 1 education participants. Yoga and PT did not differ significantly in frequency or severity of adverse events (Table 11 of Supplement 2).
In a trial of yoga, PT, and education for predominantly low-income, racially diverse participants with moderate to severe nonspecific cLBP, intention-to-treat analyses found that a 12-week standardized yoga program was noninferior to individually delivered PT for change in back-related function and pain. Our secondary hypothesis, that yoga is superior to education for both function and pain, was not supported. However, participants in both yoga and PT were more likely to have clinically meaningful improvements in function than were education participants. Yoga and PT participants were also more likely than education participants to discontinue pain medication. Improvements in yoga and PT groups were maintained at 1 year regardless of whether patients were assigned to ongoing yoga classes, PT booster sessions, or home practice only. All interventions were relatively safe.
Compared with previous trials of yoga for cLBP (11–13), our trial enrolled a more racially diverse, lower SES population. For example, Sherman and colleagues’ study of yoga, stretching classes, and education recruited 228 patients with cLBP, 87% of whom were white, 62% college graduates, and 84% with incomes greater than $45 000 (12). Research has documented racial and socioeconomic disparities in disability and pain (5). Minorities with back pain receive fewer specialty referrals (43) and less-intensive rehabilitation for occupational back injuries (44) than do whites. Despite pain’s disproportionate impact on minority and low SES groups, few cLBP studies and even fewer yoga and PT trials have targeted these populations. Barriers exist for low-income minorities to access nonpharmacologic treatments, such as yoga and PT. Non-Hispanic white adults are twice as likely as non-Hispanic black adults to use yoga (45). Yoga classes are often unavailable in predominantly low-income minority neighborhoods, and fees can be prohibitive (46). People with higher education are more likely to receive PT; people covered by Medicaid are less likely (47). Some insurance plans require expensive PT copayments that may deter patient access (48).
In Sherman and colleagues’ study, yoga classes and stretching classes were superior to the same self-care book used in our study (12). Although both our study and Sherman and colleagues’ study found similar improvement for patients who used the book (for example, −2.5 vs. −2.2 for RMDQ), our yoga intervention did not perform as well (for example, −3.8 vs. −5.2 for RMDQ). One possible explanation for the modest effect of yoga and PT in our trial is lower adherence. Our yoga participants attended a median of 7 classes compared with 10 in Sherman and colleagues’ trial. This can be attributed in part to obstacles often facing lower SES populations, including inconsistent telephone service, difficulties with transportation, serious illnesses and injuries in family members, conflicting life demands (such as work, child care, and elder care), homelessness, and incarceration (49). The greater effect of yoga and PT among our adherent patients highlights the challenges and importance of compliance in exercise interventions (50). Another reason for the modest effect of our interventions may be the severity of our participants’ back conditions. Baseline mean back-related disability and pain scores were 63% and 57% more severe, respectively, in our study than in Sherman and colleagues’ study. Opioid use, obesity, depression, and other comorbid conditions were also more common.
Strengths of this study include an assessor-blinded randomized design, adequate power to assess noninferiority of yoga to PT at 12 weeks, and standardized interventions delivered by providers in community-based settings. Although blinding of participants in nonpharmacologic trials is not possible, we presented the study as comparing 3 credible treatments. Limitations include disproportionate loss to follow-up for PT. This attrition bias could increase or decrease the observed effectiveness of PT depending on the likelihood of dropouts doing worse or better than those with follow-up. Mean baseline RMDQ, a coprimary outcome, modestly differed between groups and required adjustment. The lack of difference between maintenance phase groups is difficult to interpret because only approximately half of eligible participants attended any yoga drop-in classes or PT booster sessions. Per protocol analyses were not powered to assess noninferiority and should be interpreted with caution. Because the per-protocol population is a subgroup, differences between adherent and nonadherent participants may confound the analyses. In addition, analyses for change in use of medication subtypes, such as opioids or nonsteroidal anti-inflammatory drugs, were underpowered.
These findings suggest that a manualized yoga intervention designed specifically for cLBP is similarly effective to PT for improving physical function and reducing pain in a diverse underserved population with high levels of impairment. These results are likely generalizable to other safety-net settings and to less-impaired, higher SES patients. However, the results may not generalize to typical nonmanualized, community-based yoga classes. Similarly, our education intervention was more time-intensive than typical office-based education and therefore was probably more effective. Future studies should focus on pragmatic trials of nonmanualized yoga classes, testing of strategies to enhance adherence, and cost-effectiveness analyses.
In conclusion, we found that yoga was noninferior to PT for improving moderate to severe nonspecific cLBP in a diverse, predominantly low-income population. Yoga and PT participants had greater improvement in function and pain than education participants; however, these differences were not uniformly significant. Yoga and PT participants were more likely than education participants to stop taking pain medication. The effectiveness of yoga and PT was most evident in adherent participants. Improvements in yoga and PT were maintained at 1 year, and all interventions were relatively safe. A structured yoga program for patients with cLBP may be a reasonable alternative to PT depending on patient preferences, availability, and cost.